Roadside vehicle detection system provides critical data support for intelligent transportation systems. However, traditional vehicle detection technologies still face challenges such as high maintenance and deployment costs. The Distributed Acoustic Sensing (DAS) system has been shown to offer passive, wide-range, and high spatial resolution vehicle detection and localization, without the need for on-site sensor installation, thus providing a significant advantage in terms of ultra-low deployment costs. However, DAS-based vehicle detection systems are susceptible to environmental noise and system fading. To address this issue, this paper proposes a Multi-Scale Context Generative Adversarial Network (MCGAN), which extracts features using multi-scale dilated convolutions and integrates hierarchical information across different scales to enhance the model's denoising performance. Experimental results demonstrate that MCGAN can effectively suppresses environmental and fading noise while preserving vehicle signals, particularly excelling in detecting low-speed and small vehicle signals.